The v8a feature should be enabled by default on aarch64

[adamgemmell]

The TTBR0_EL2 register is locked behind el2vmsa when it should be available on all aarch64 platforms. This has been noticed in Rust which does not currently enable the feature by default or expose a stable way to enable it, though I feel here is a better place to solve the issue.

The v8a feature, added about a year ago, seems to want to represent features enabled on ARM v8.0 processors, but these features should really be enabled by default if this is the case.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

[oToToT]

I believe this issue is similar to #53956.

LLVM currently takes the intersection of v8a and v8r as default. 4a31af8 tried to provide some backward compatibility (i.e. v8a as default) after 8689f5e landed.
However, the use case of Rust is not covered in that patch (which I misunderstood then), so either we need to figure out a better solution to provide backward compatibility or we could simply modify the target spec of Rust.

[lenary]

just for context: the AArch64 backend now supports both v8-r and v8-a, which have significant overlap. For this reason, the backend cannot always assume that it has v8-a, because you might be compiling for v8-r. The same goes for disassembly.

In fact, disassembly is harder because v8-a and v8-r have some conflicts in their system register name/encoding space, so you have to know if you're aiming for r-profile or a-profile when disassembling. We recently added functionality to objdump to turn on all features, but this was probably incorrectly designed because it does indeed switch on everything, across both A and R profile, even though there are conflicts.

[adamgemmell]

Thanks @oToToT!